1 IDENTIFYING ACTIVE MEMBERS IN VIRAL MARKETING CAMPAIGNS

Olivier Toubia, Andrew T. Stephen, and Aliza Freud

November 26, 2010

Draft – Please do not cite without permission.

Olivier Toubia is Associate Professor of Business, Columbia Business School ([email protected]); Andrew T. Stephen is an Assistant Professor of Marketing, INSEAD ([email protected]); Aliza Freud is founder and CEO, SheSpeaksTM ([email protected]). This paper has benefited from the comments of the participants in the BRITE conference on online communities at Columbia Business School and the “Practice and Impact of Marketing Science” conference at the MIT Sloan School of Management. We wish to thank Oded Netzer and Avi Goldfarb for their useful suggestions.

2 IDENTIFYING ACTIVE MEMBERS IN VIRAL MARKETING CAMPAIGNS

ABSTRACT Viral marketing campaigns (VMCs) are promotional tools whereby companies seed products with select groups of consumers in the hope that these consumers will spread information and marketing materials about these products. Despite the growing popularity of VMCs, it is not yet fully understood which consumers a company should target when planning a VMC. Information on at least two types of consumer characteristics are often available to managers: (i) product-related characteristics that capture a consumer’s relation to the product, brand, and category, and (ii) social characteristics that measure a consumer’s social connectivity and perceived social influence. Using data from two real-world VMCs covering different product categories we identify which of these consumer characteristics are associated with greater levels of activity. We find that social characteristics—which are not campaign-specific—are better predictors of level of activity compared to product-related consumer characteristics. Moreover, social characteristics may be used to identify some of the more active members in a campaign before the start of the campaign.

3 INTRODUCTION Viral marketing has become an increasingly popular promotions and marketing communications tool (Kirsner 2005; Walker 2004). According to a 2009 study by the media research firm PQ Media, spending on word-of-mouth (WOM) marketing in the US rose at a compound annual growth rate of 53.7% from 2001 to 2008—from US$76 million to US$1,543 million—and is forecast to reach over US$3 billion annually by 2013 (PQ Media 2009). However, in a 2007 survey of marketing and advertising professionals by Dynamic Logic, half the respondents rated viral marketing as more of a fad than a mainstream and widely available tactic (Dynamic Logic 2007). One of the challenges currently faced by the viral marketing industry is the lack of systematic methods for optimizing viral marketing campaigns (VMCs). We refer to viral marketing (also sometimes called word-of-mouth marketing, buzz marketing, or social engagement marketing) as a set of promotional tools whereby companies seed products with select groups of consumers (commonly called “seeds”) in the hope that they will spread information and marketing materials about these products. Firms offering such services include BzzAgentTM (www.bzzagent.com), SheSpeaksTM (www.shespeaks.com), Procter and Gamble’s VocalpointTM (www.vocalpoint.com), and Vocanic (www.vocanic.com). Each of these firms has built online panels of consumers from whom campaign seeds are selected. These seeds are (i) offered early access to new products (e.g., by being sent free samples), (ii) encouraged to share information and marketing materials about these products with other consumers (e.g., promotional offers such as coupons), (iii) asked to participate in surveys to provide feedback on the products to marketers, and (iv) asked to report their sharing activities to the firm on a regular basis.

4 Recently, academic researchers have started to examine viral marketing as a new tool for marketing communication and promotion. Godes and Mayzlin (2009) use a largescale field test to study word-of-mouth (WOM) transmissions in VMCs, in particular how the effectiveness of WOM in a VMC varies as a function of whether the originator of WOM is a loyal customer of the brand being promoted, and of whether WOM is targeted to a friend or an acquaintance. De Bruyn and Lilien (2008) develop a model to identify the role that WOM plays during each stage of a viral marketing recipient’s decisionmaking process. Biyalogorsky, Gerstner, and Libai (2001) study customer referral programs theoretically and identify conditions under which they should be used. Van der Lans, van Bruggen, Eliashberg and Wierenga (2010) propose a branching model for predicting the viral spread of online marketing content. Iyengar, Van den Bulte and Valente (2010) study how the degree to which a customer influences other customers varies with his or her position in their social network. Trusov, Bucklin and Pauwels (2009) show that WOM referrals have a greater effect on the growth of an Internet social networking site, as compared to traditional marketing efforts. Berger and Schwartz (2010) study the effect of product characteristics on the amount of WOM generated in a VMC, finding that the most buzzed-about products are not necessarily the most interesting or exciting. The present paper focuses on identifying some characteristics of the more active consumers in VMCs. In our setting, the level of activity of a consumer in a VMC is measured by the number of coupons shared by this consumer with other consumers. We use data from two successive real-world campaigns in different product categories. The availability of data from two consecutive campaigns allows us not only to test whether

5 the findings from the first campaign are replicated in the second, but also to assess whether some of the more active members in the second campaign may be identified before the start of that campaign, based on a model calibrated on the data from the first campaign. Our research provides both conceptual and practical contributions. Conceptually, previous research on WOM and social interactions between consumers has studied the impact of the characteristics of the dyad formed by the source and the recipient of information (e.g., tie strength, homophily) on the amount of WOM generated (Brown and Reingen 1987). Researchers have also studied the impact of these dyad characteristics as well as characteristics of the source and the recipient themselves (e.g., expertise, opinion leadership) on the propensity of the recipient to actively seek information over social ties (Bansal and Voyer 2000; Brown and Reingen 1987) and on the impact that WOM has on the recipient’s behavior (Bansal and Voyer 2000; Brown and Reingen 1987; Gilly et al. 1998). Other research (reviewed below) has suggested that more social consumers (e.g., those with many social ties) are likely to be more active if selected as seeds in VMCs. However, it is not yet known how well these social seed characteristics predict activity compared to other seed characteristics routinely measured in the context of VMCs, and in particular those characteristics that describe the relation between the seed and the product being promoted (e.g., attitude toward the product, brand, and category). In fact, it is common practice in the viral marketing industry to use these product-related consumer characteristics as a basis for seed selection. Therefore, the present research may help managers of VMCs optimize the selection of seeds by suggesting the types of seed characteristics to pay more attention to when screening potential campaign participants.

6 The paper is organized as follows. We first describe some theoretically relevant seed characteristics. We then describe our data, and formulate a series of alternative models to capture the impact of these characteristics on seeds’ levels of activity in VMCs. We then present our results, and finally conclude with a discussion of implications and directions for future work in this area.

PRODUCT-RELATED AND SOCIAL SEED CHARACTERISTICS Viral marketing companies usually collect data on at least two types of characteristics of their panel members who could be used as seeds: (i) product-related characteristics that capture the member’s attitudes towards the product, brand, and category involved in each campaign, and (ii) social characteristics that measure the member’s social connectivity and perceived social influence. One important distinction between these two types of characteristics is that product-related seed characteristics are campaign-specific while social seed characteristics are not.1 Therefore, whether product-related seed characteristics are required to identify more active members of a given viral marketing campaign affects a marketer’s ability to identify and target a set of active seeds before the start of a campaign. While one should expect seeds who possess certain product-related characteristics (e.g., having a high level of knowledge of or familiarity with the product category, being a strong advocate of the focal brand) to be more active in a VMC, the social characteristics of the seeds are also likely to be important. Indeed, because the success of a VMC hinges on the generation of social interactions between seeds and other 1

Social characteristics may also be measured in a campaign-specific way; e.g., by asking members how many people they talk to on a daily basis about a specific product category. We leave the exploration of the efficacy of such measures to future research.

7 consumers, and because the propensity to engage in social interactions is a function of a seeds’ social characteristics, such characteristics should also be associated with higher levels of activity. We focus on three main social characteristics, which are grounded in previous research. The first two are measures of a member’s social connectivity, and the third measures the self-perceived impact of his or her recommendations. The first two characteristics, which capture the number of social connections a member has and how “social” he or she is, were part of the “network breadth” scale proposed in an earlier version of Godes and Mayzlin (2009).2 Godes and Mayzlin found some preliminary yet inconclusive evidence that members who score high on that scale are more active in VMCs. The rationale behind these items is that, all else equal, the number of social interactions about a specific product is increasing in the overall total number of social interactions in which a member typically engages, and how generally “social” that person is. Some justification for this argument may be traced back to Coleman, Katz, and Menzel (1957), who show that physicians with high connectivity (degree centrality) in their social networks play a critical role in diffusion (although this finding has been brought into question by Van den Bulte and Lilien 2001). Similarly, Iyengar, Van den Bulte and Valente (2010), in their study of the diffusion of a new prescription drug among physicians, find that a physician with a larger in-degree centrality (e.g., number of other physicians who nominated that physician as someone with whom they felt comfortable discussing the clinical management and treatment of the disease) has a greater impact on the adoption of other physicians. Goldenberg et al. (2009) empirically show that individuals who are highly central in a social network (i.e., 2

This scale is not mentioned in the final published version of Godes and Mayzlin (2009).

8 “hubs” who have an exceptionally large number of connections) are critical for accelerating the diffusion of an innovation over this network. The third social characteristic that we consider captures the perceived effect of these social interactions on recipients’ behaviors (e.g., “When you recommend products or services to friends, what do they usually do?”). The rationale behind this item is that members who believe that their recommendations have more impact should be more likely to spread information about a product. Indeed, Stephen and Lehmann (2010) show experimentally that consumers take into account how likely they believe a given receiver will be to listen to them before deciding whether or not to share product-related information with that person. The social and product-related member characteristics collected in the first VMC analyzed in this paper are reported in Table 1 (similar data were collected in the second VMC, but the identity of the product in that campaign is confidential). These data are collected for every campaign run by SheSpeaksTM, the viral marketing firm with which we collaborated. Similar data are collected by other viral marketing firms (see for example Godes and Mayzlin 2009). [INSERT TABLE 1 ABOUT HERE] Note that our measures do not include the standard opinion leadership scale (e.g., Rogers and Cartano 1962; King and Summers 1970). The role of this scale in viral marketing was studied by Godes and Mayzlin (2009). Their field experiment involved a set of “agents” from BzzAgent (a viral marketing firm), and a set of customers enrolled in the company’s loyalty program. They found that while opinion leadership had an impact on WOM activity among members of the company’s loyalty program, it had no impact

9 among the BzzAgents.

DATA Our data come from two VMCs conducted by SheSpeaksTM, a firm specialized in running VMCs targeted to women. The first campaign was conducted in collaboration with OPI, a leading manufacturer of cosmetic products, in the Spring/Summer of 2008 as the company was launching a new product, Nic’s Sticks. Nic’s Sticks were innovative nail polish pens designed for easy and quick application. The polish was applied using a brush at one end of the pen. A clickable pump at the other end allowed the user to control the flow of polish (see Figure 1). The product was available in a wide range of colors, creating opportunities for multiple purchases per consumer. The second campaign was for a packaged food product, and ran several months after the OPI campaign.3 As will be seen later, there was very little overlap in the sets of members involved in the two campaigns. [INSERT FIGURE 1 ABOUT HERE] Both campaigns followed the typical steps of a VMC. First, a random subset of the SheSpeaks™ panel was invited to participate in the campaign by email. Second, members filled out an “enrollment” survey that assessed their initial dispositions towards the category, the brand, and the specific product targeted in the campaign product. We obtained our measures of product-related seed characteristics from this survey. These characteristics are described in the bottom half of Table 1. While such measures are typically used for seed selection by SheSpeaksTM (by picking the members

3

For reasons of confidentiality the brand name and specific product details for Campaign 2 cannot be disclosed.

10 who are most experienced with the brand and/or the product category), members were not screened in these two VMCs in order to avoid any selection bias. Instead, members who responded to the initial invitation email and filled out the enrollment survey were randomly chosen to participate in the campaign as seeds. Third, the selected seeds were mailed a package containing a free sample of the product, five coupons (see Figure 1 for the coupon used in Campaign 1), and a doublesided postcard introducing the product and the campaign. Fourth, several weeks later, seeds were invited by email to complete an “evaluation” survey asking for their post-usage product evaluations. Each seed was asked to report the number of coupons that she used to buy the product for herself as well as the number of coupons that she shared with others (i.e., used to buy a product for others or given to others).4 We use the number of coupons shared with others as the measure of the seed’s level of activity in the campaign. This metric is critical to the VMC industry, as it is the basis for measures of return on investment used to evaluate campaign efficiency. Therefore, being able to identify seeds who will be more active in sharing coupons with others is of high importance to both viral marketing firms and their clients. In addition to these steps, all members filled out an initial survey when they joined the panel (i.e., before participating in their first campaign). Data on social seed characteristics as well as demographic data come from this survey, which is independent of any campaign. The social seed characteristics data are described in the top half of Table 1. The demographic data consist of age and employment status. Age was measured using seven categories: below 24, 25-29, 30-34, 35-39, 40-45, 45-49, 50 or above.

4

We removed from the analysis any member who claimed to have used more than five coupons, since only five coupons were sent to each member.

11 Employment status was measured with a binary variable indicating whether or not the member was employed full-time.

MODELING THE IMPACT OF SEED CHARACTERISTICS ON ACTIVITY As mentioned above, the level of activity of a seed is measured as the number of coupons she shared with other consumers. Because coupons can be shared with others or used by the seed to buy products for herself, we need to jointly model these two outcome measures. We denote nself,i and nothers,i as the number of coupons used by seed i for herself and shared with others, respectively. Because five coupons were sent to each seed, the following constraints hold: nself ,i ≥ 0, nothers,i ≥ 0, and nself ,i + nothers,i ≤ 5. The existence of two interconnected and jointly constrained dependent variables precludes the use of traditional regression techniques. Also, neither extant literature nor practical considerations suggest a single best and most appropriate model of activity in this context. In particular, it is not clear whether a seed jointly decides how to allocate her set of coupons between herself and others (i.e., a single joint allocation decision) or undergoes a two-step decision process where she first decides how many coupons to use for herself and then decides how many of the remaining coupons to share with others, or vice versa. Also, in the case of a two-step process it is not clear whether a seed myopically maximizes her utility or adopts a more forward-looking perspective. In the absence of theory or practical directions suggesting a single most appropriate model, we test the robustness of our findings under a wide range of model assumptions. We start with a base model that assumes seeds decide jointly how many coupons to share with others and to use for themselves. We then develop and estimate

12 two alternative dynamic discrete choice models (Erdem and Keane 1996; Imai, Jain and Ching 2009; Rust 1987) that assume a sequential decision process (self-then-others and others-then-self). These models are estimated across the full range of possible discount factors (i.e., the degree to which seeds are forward-looking in their coupon allocation decisions), from myopic to joint allocation. The models are presented in this section and estimation results are presented in the next section.

Utility Specification We model the utility derived by seed i from choosing to allocate nself,i coupons to herself as: u1,i (nself ,i ) = uself ,i (nself ,i ) ! cself ,i (nself ,i ) + !1 (nself ,i ) , where uself ,i (nself ,i ) captures the utility derived from using nself,i coupons for herself, cself ,i (nself ,i ) captures the corresponding transaction costs, and the term !1 (nself ,i ) is an unobservable component of utility that is known to the seed but not the researcher. Similarly, we model the utility derived by seed i from deciding to share nothers,i coupons with others as:

u2,i (nothers,i ) = uothers,i (nothers,i ) ! cothers,i (nothers,i ) + ! 2 (nothers,i ) , where uothers,i (nothers,i ) captures the utility derived from sharing nothers,i coupons with others, cothers,i (nothers,i ) captures the corresponding transaction costs, and ! 2 (nothers,i ) is an unobservable component of utility. We denote as Xi the row vector containing the set of covariates describing seed i. Xi is composed on an intercept and covariates that capture the demographic, social, and product-related seed characteristics for seed i. We define as βself and βothers vectors (which we estimate) that capture the relation between the seed characteristics captured by Xi and the utility from using one coupon for oneself and from sharing one coupon with others,

13 respectively. In particular, uself ,i (1) = X i β self and uothers,i (1) = X i βothers . The marginal utility j −1 from the jth coupon used for oneself and shared with others are modeled as γ self ⋅ X i β self

j −1 and γ others ⋅ X i β others respectively, where the parameters γself ≥ 0 and γothers ≥ 0 capture

variations in marginal utilities from additional coupons. This gives rise to the following utility specification:

$ 0 & nself ,i j!1 uself ,i (nself ,i ) = % &' (" j=1 ! self )# (Xi # " self ) $ 0 & n uothers,i (nothers,i ) = % others,i j!1 &' (" j=1 ! others )# (Xi # "others )

if nself ,i = 0 if nself ,i > 0

if nothers,i = 0 if nothers,i > 0

(1)

(2)

We model the transaction costs cself ,i (nself ,i ) and cothers,i (nothers,i ) as being proportional to the number of coupons: cself , i (nself ,i ) = nself , i ⋅ θ self and

cothers,i (nothers,i ) = nothers,i ! ! others , where θself and θothers are two non-negative parameters (which we estimate) that capture the unit transaction costs for using one coupon for oneself and for sharing one coupon with others, respectively.

Base Model: Joint Allocation of Coupons Our base model assumes that seed i jointly selects nself,i and nothers,i by solving the following single period optimization problem: {nself ,i , nothers,i } = argmax[uself ,i (nself ) − cself ,i (nself ) + ε1 (nself ) + uothers,i (nothers ) − cothers,i (nothers ) + ε 2 (nothers )] nself ,nothers

subject to : 0 ≤ nothers, ; 0 ≤ nself ; nself + nothers ≤ 5

Under the standard assumption that ε follows a double exponential distribution,

14 the likelihood function is: Pr({nself ,i , nothers,i }) =

exp[uothers,i (nothers,i ) ! cothers,i (nothers,i ) + uself ,i (nself ,i ) ! cself ,i (nself ,i )] 5

5!nself

" "

(3)

exp[uself ,i (nself ) ! cself ,i (nself ) + uothers,i (nothers ) ! cothers,i (nothers )])

nself =0 nothers =0

Two Alternative Dynamic Models: Sequential Allocation of Coupons We also develop two sequential models: one assuming the seed first allocates coupons to herself and then to others, and the other assuming she first allocates coupons to others and then to herself. One allocation decision is made per period. By varying the discount factor between the two periods (i.e., how much a seed discounts period 2 utility when making her decision in period 1), each of these models allows for both myopic and forward-looking behavior. Moreover, we show in Appendix 2 that each dynamic model accepts the base model as a special case when there is no discounting. For ease of exposition, we focus here on the model that assumes seeds allocate coupons to themselves (nself,i) in Period 1 and to others (nothers,i) in Period 2 (i.e., self-then-others). The other model (others-then-self) is presented in Appendix 1. This decision making process is captured by a two-period Dynamic Program. Formally, seed i is assumed to solve the following optimization problems in sequence: Decision 1 : nself ,i = argmax[uself ,i ( nself ) − cself ,i ( nself ) + ε1 ( nself ) + δE [V2 ( nself , ε 2 )]] nself

ε2

subject to : 0 ≤ nself ≤ 5 Decision 2 : nothers,i = argmax[uothers,i ( nothers ) − cothers,i ( nothers ) + ε 2 ( nothers )] nothers

subject to : 0 ≤ nothers ≤ 5 − nself ,i

Where δ is a discount factor, and

15

V2 (nself , ε 2 ) =

max

nothers | 0 ≤ nothers ≤ 5 − n self

[uothers,i (nothers ) − cothers,i (nothers ) + ε 2 (nnothers )] is the value

function for seed i after selecting nself in period 1 and given ε2. Under the Conditional Independence Assumption typically used in dynamic discrete choice modeling and under the other typical assumption that ε1 and ε2 follow a double exponential distribution (Rust 1987), this problem may be re-written as: Decision 1 : nself ,i = argmax[uself ,i ( nself ) − cself ,i ( nself ) + ε1 ( nself ) + δV2 ( nself )] nself ,i

subject to : 0 ≤ nself ≤ 5 Decision 2 : nothers,i = argmax[uothers,i ( nothers ) − cothers,i ( nothers ) + ε 2 ( nothers )] nothers ,i

subject to : 0 ≤ nothers ≤ 5 − nself ,i 5−nself

where V2 (nself ) = log(

∑ exp[u

others,i

(nothers ) − cothers,i (nothers )])

nothers =0

and the likelihood function is then given by: Pr({nself ,i , nothers,i }) = P(nothers,i | nself ,i )! P(nself ,i ) = exp[uothers,i (nothers,i ) " cothers,i (nothers,i )]

'

(4)

!

5"nself ,i

exp #$uothers,i (nothers ) " cothers,i (nothers )%&

nothers =0 5"nselfi

exp[uself ,i (nself ,i ) " cself ,i (nself ,i ) + ! log(

'

exp[uothers,i (nothers ) " cothers,i (nothers )])]

nothers =0 5"nself # % ( ) exp u (n ) " c (n ) + ! log( exp [u (n ) " c (n )]) ' ( self ,i self self ,i self ' others,i others others,i others )& n =0 $ nself =0 others 5

The myopic case in which seed i does not take into account any future utility from sharing coupons when deciding how many coupons to use for herself is obtained as a special case with δ = 0. In this case the likelihood function becomes:

16

Pr({nself ,i , nothers,i }) = P(nothers,i | nself ,i )! P(nself ,i ) = exp[uothers,i (nothers,i ) " cothers,i (nothers,i )] 5"nself ,i

'

exp #$uothers,i (nothers ) " cothers,i (nothers )%&

nothers =0

!

exp[uself ,i (nself ,i ) " cself ,i (nself ,i )] 5

'

(5)

exp #$uself ,i (nself ) " cself ,i (nself )%&

nself =0

RESULTS We first analyze the base model and then investigate whether the results are robust to the two dynamic models estimated under the full range of discount factors. We estimate the parameters βself, βothers, γself, γothers, θself, and θothers using maximum likelihood.5 The N

likelihood function is L = ! Pr({nself ,i , nothers,i }) . Our analysis is based on the sample of i=1

seeded SheSpeaks™ members for whom all data were available, resulting in 1,032 complete observations in Campaign 1 and 1,597 in Campaign 2. Only 71 SheSpeaks™ members were seeds in both campaigns, therefore there is only very little overlap in our data between participants from both campaigns.

Campaign 1 Main findings. We report parameter estimates for the base model in Table 2. The analysis of the parameter estimates is not straightforward, for at least two reasons. First, both sets of parameters βself and βothers influence both the number of coupons used for oneself and the number of coupons shared with others. Second, βself and βothers in the model only capture the utility from one coupon. The total number of coupons used for oneself and shared with others also depends on the other parameters of the model. Moreover, while 5

Because the state space in our dynamic discrete choice models is relatively small, they do not suffer from the curse of dimensionality typical in such models, and standard maximum-likelihood procedures may be used.

17 the parameter estimates per se are of some interest to this research, our real focus is on identifying active members in a VMC based on observable characteristics. [INSERT TABLE 2 ABOUT HERE] Therefore, while we report the parameter estimates for completeness in Table 2, we interpret these parameters using more intuitive measures and analyses. Let X isocial be the vector obtained by setting all the elements of Xi that capture product-related seed characteristics to 0 (i.e., keeping only the demographic and social seed characteristics). Similarly, let X iproduct be the vector obtained by setting all the elements of Xi that capture social seed characteristics to 0. We assess the impact of social and product-related seed characteristics on sharing coupons with others by computing a social score and a product score for each member, scoreisocial = X isocial . βothers and scoreiproduct = X iproduct . βothers . For completeness we also consider scoreisocial + product = X i .β others . We group seeds into four quartiles based on score isocial and compute the average number of coupons shared with others by each quartile. As the score increases (i.e., going from the 1st to 2nd quartile, 2nd to 3rd, and 3rd to 4th), the number of coupons shared with others should also increase if the information on which that score is based is informative. We do the same using score iproduct and scoreisocial+product . We report the average number of coupons shared with others by each of the quartiles in Table 3. The comparison of the quartiles based on scoresocial to those based on scoreproduct suggests that social seed characteristics are more useful in identifying groups of seeds with higher activity compared to product-related seed characteristics. The comparison of the quartiles based on scoresocial to those based on scoreproduct+social suggests that product-related seed

18 characteristics do not add much information above and beyond social seed characteristics. For each type of score we compare the average number of coupons shared by seeds in the top quartile with the average number of coupons shared by all seeds. We define the improvement in seeds’ campaign activity based on the social score (and similarly for

score iproduct and score iproduct + social ) as: improvementsocial =

Mean nothers,i for top quartile based on score isocial

(6)

– Mean nothers,i for all seeds These improvement measures indicate how much better (or worse) the expected sharing activity is when using seeds who are very high on a particular score versus using a random subset of seeds. Improvementsocial is .60, improvementproduct is .26, and improvementsocial+product is .63. The ratio of improvementsocial to improvementproduct equals 2.30; i.e., the improvement based on social seed characteristics is more than double the improvement based on product-related seed characteristics. The ratio of improvementsocial+product to improvementsocial is equal to 1.05; confirming that productrelated seed characteristics do not add much information above and beyond social seed characteristics. [INSERT TABLE 3 ABOUT HERE] Robustness to model specification. We also estimate the two dynamic discrete choice models described in the previous section and Appendix 1 (i.e., self-then-others, and others-then-self). We use values for the discount factor (δ) that cover the full possible range since our data do not allow us to estimate this parameter. Recall that the coupon allocation decision is myopic when δ = 0, and joint when δ = 1 (see Appendix 2). We

19 estimate both models for δ = {0, .1, .2, .3, .4, .5, .6, .7, .8, .9, .95, .99, 1}.6 For each model and each value of δ, we compute the ratios improvementsocial / improvementproduct and improvementsocial+product / improvementsocial. These ratios are plotted in Figure 2 for the range of discount factors. The ratio of improvementsocial / improvementproduct (top panel in Figure 2) lies between 2.30 and 2.67 across the levels of δ under the self-then-others model, and between 1.80 and 2.42 under the others-then-self model. The ratio improvementsocial+product / improvementsocial (bottom panel in Figure 3) ranges between 1.02 and 1.10, and .88 to 1.05 for the self-then-others and others-then-self models, respectively. These results demonstrate the robustness of the findings to model specification; i.e., irrespective of the assumed behavioral model of coupon usage and discount factor, using social seed characteristics to identify more active seeds is more efficient than using product-related seed characteristics, and using both sets of characteristics together adds little (or hurts slightly) over using the social characteristics only. [INSERT FIGURE 2 ABOUT HERE]

Campaign 2 The same analysis procedure used for campaign 1 was followed for campaign 2. We report parameter estimates for the base model estimated on the campaign 2’s data in Table 4, and report mean numbers of coupons shared with others across quartiles based on various scores in Table 5. The improvement measures and the ratios of

6

Note that our goal is not to find the discount factor that maximizes fit, but rather to check that our main findings are robust to a range of assumptions regarding (i) the underlying decision model used by seeds when allocating coupons between themselves and others, and (ii) how myopic or forward-looking seeds are when making these decisions.

20 improvementsocial to improvementproduct and improvementsocial+product to improvementsocial were also computed as before. The ratio of improvementsocial to improvementproduct is 1.45 and the ratio of improvementsocial+product to improvementsocial is 1.04. The general findings are consistent with those reported for campaign 1; i.e., compared to product-related seed characteristics, social seed characteristics are more useful in identifying groups of seeds with higher activity, and product-related seed characteristics do not add much information above and beyond social seed characteristics. [INSERT TABLES 4 AND 5 ABOUT HERE] Finally, we perform the same robustness analyses that we did for campaign 1 by estimating the two dynamic discrete choice models under the full range of discount factors. In Figure 3 we plot the ratios of improvementsocial to improvementproduct (top panel) and improvementsocial+product to improvementsocial (bottom panel) over the range of discount factors. Our findings appear to be robust to different model specifications. [INSERT FIGURE 3 ABOUT HERE]

Out-of-Sample Prediction: Using Campaign 1 to Predict Campaign 2 Activity Since social seed characteristics appear to be better than product-related seed characteristics as predictors of activity, and social characteristics are seed-specific measures independent of a particular campaign, we tested how well they could be used to identify groups of more active seeds before the start of a campaign. Specifically, we tested whether the social seed characteristics measured when seeds joined the SheSpeaksTM panel can be used to identify some of the more active members in Campaign 2 before Campaign 2 starts, i.e., without using any data from Campaign 2. For

21 this we computed a social score for each Campaign 2 seed based on the estimate of βothers from Campaign 1. Importantly, this score does not use any data from Campaign 2. We then performed the quartile analysis described above. The mean numbers of coupons shared with others in Campaign 2 for the social score quartiles based on Campaign 1 estimates are included in Table 5 for the base model, and the ratio of improvementsocial to improvementproduct across models and values of the discount factor is reported in Figure 4. The results suggest that our ability to identify more active seeds using campaign-independent social covariates only is equally good when no data from Campaign 2 are used. Interestingly, the out-of-sample predictive ability of social seed characteristics is greater than the in-sample predictive ability of product-related seed characteristics; i.e., social characteristics combined with parameter estimates from a previous campaign are more informative than campaign-specific product-related seed characteristics combined with data from the target campaign. [INSERT FIGURE 4 ABOUT HERE]

DISCUSSION AND CONCLUSION To summarize, the results of our analyses of two different VMCs suggest that (i) compared to product-related seed characteristics, social seed characteristics are more useful for identifying groups of seeds with higher activity, (ii) product-related seed characteristics do not add much information above and beyond social seed characteristics, and (iii) some of the more active seeds in VMCs may be identified before the start of a campaign using data from a prior campaign and campaign-independent social seed characteristics, without requiring any data from the target campaign. These results are

22 robust to various model specifications that assume a joint allocation of coupons or a sequential allocation with varying sequences and discount factors. While we have focused on coupon sharing as the desired activity in the kinds of VMCs we examine, it is important to remember that sharing a coupon with someone else in most cases will also involve having a conversation about the product associated with the coupon. Thus, seeds who are more active in sharing coupons should also be more likely to generate product-related WOM conversations. Our research is not without limitations. In particular, the data used from both campaigns were self-reported. Coupons are usually not tracked individually in these campaigns, and therefore it was not possible to obtain individual-level redemption data without relying on self-reports. Data on social and product-related seed characteristics were also self-reported and therefore subject to the same caveats. While self-reported data have some shortcomings, they are commonly used in the viral marketing industry and literature (see also Berger and Schwartz 2010; Godes and Mayzlin 2009). Future research may explore other sources of data on social seed characteristics (e.g., number of Facebook friends or Twitter followers). Also, while our findings are robust across two campaigns, we make no claims of generalizability, and we encourage future research that addresses similar and related questions. Finally, another interesting area for future research is comparing the effectiveness of VMCs to that of other “traditional” advertising such as television and print advertising.

23 REFERENCES Bansal, Harvir S., and Peter A. Voyer (2000), “Word-of-Mouth Processes Within a Services Purchase Decision Context,” Journal of Service Research, 3(2), 166-177. Biyalogorsky, Eyal, Eitan Gerstner and Barak Libai (2001), “Customer Referral Management: Optimal Reward Programs,” Marketing Science, 20(1), 82-95. Berger, Jonah, and Eric Schwartz (2010), “What Products Get Talked About? How Cues, Public Visibility, and Interest Shape Immediate and Ongoing Word-of-Mouth,” working paper, Wharton School. Brown, Jacqueline Johnson, and Peter H, Reingen (1987), “Social Ties and Word-ofMouth Referral Behavior,” Journal of Consumer Research, 14 (December), 350362. Coleman, James, Elihu Katz, and Herbert Menzel (1957), “The Diffusion of an Innovation Among Physicians,” Sociometry, 20 (4), 253-270. De Bruyn, Arnaud, and Gary L. Lilien (2008), “A Multi-Stage Model of Word-of-Mouth Influence Through Viral Marketing,” International Journal of Research in Marketing, 25, 151-163. Dynamic Logic (2007), “Many Marketers Think Viral is a Fad,” http://www.dynamiclogic.com/na/pressroom/coverage/?id=491, Mar. 22. Erdem, Tülin, and Michael P. Keane (1996), “Decision-Making under Uncertainty: Capturing Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets,” Marketing Science, 15(1), 1-20.

Gilly, Mary C., John L. Graham, Mary Finley Wolfinbarger, and Laura J. Yale (1998), “A Dyadic Study of Interpersonal Information Search,” Journal of the Academy of Marketing Science, 26(2), 83-100.

24 Godes, David and Dina Mayzlin (2009), “Firm-Created Word-of-Mouth Communication: Evidence from a Field Test,” Marketing Science, 28(4), 721-739. Goldenberg, Jacob, Sangman Han, Donald R. Lehmann, and Jae Weon Hong (2009), “The Role of Hubs in the Adoption Process,” Journal of Marketing, 73 (2), 1-13. Imai, Susumu, Neelam Jain, and Andrew Ching (2009), “Bayesian Estimation of Dynamic Discrete Choice Models,” Econometrica (77), 1865–1899. Iyengar, Raghuram, Christophe Van den Bulte, and Thomas Valente (2010), “Opinion Leaderhip and Social Contagion in New Product Diffusion,” forthcoming, Marketing Science. King, Charles W., and John O. Summers (1970), “Overlap of Opinion Leadership Across Consumer Categories,” Journal of Marketing Research, 7(1), 43-50. Kirsner, Scott (2005), “How Much Can You Trust,” The Boston Globe, 11/24/2005. PQ Media (2009), “Word-of-Mouth Marketing Forecast 2009-2013: Spending, Trends & Analysis,” Stamford, CT: PQ Media. Rogers, Everett M., and David G. Cartano (1962), “Methods of Determining Opinion Leadership,” Public Opinion Quarterly, 26(3), 435-441. Rust, John (1987), “Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold Zurcher,” Econometrica, 55(5), 999-1033. Stephen, Andrew T. and Donald R. Lehmann (2010), “To Whom Do Consumers Transmit Organic Word-of-Mouth?” working paper, INSEAD. Trusov, Michael, Randolph E. Bucklin and Koen Pauwels (2009), “Effects of Word-ofMouth Versus Traditional Marketing: Findings from an Internet Social Networking Site,” Journal of Marketing, 73 (September), 90-102.

25 Van den Bulte, Christophe and Gary L. Lilien (2001), “Medical Innovation Revisited: Social Contagion versus Marketing Effort,” American Journal of Sociology, 106 (5), 1409-1435. Van der Lans, Ralf, G. H. van Bruggen, Jehoshua Eliashberg, and B. Wierenga (2009), “A Viral Branching Model for Predicting the Spread of Electronic Word-ofMouth,” Marketing Science, forthcoming. Walker, Rob (2004), “The Hidden (In Plain Sight) Persuaders,” The New York Times, 12/04/2004.

26 FIGURE 1 PRODUCT AND COUPON USED IN CAMPAIGN 1

27 FIGURE 2 ROBUSTNESS OF THE RESULTS TO MODEL SPECIFICATION – CAMPAIGN 1

3.00

Ratio of improvementsocial to improvementproduct

2.50

2.00

1.50 Self first, Others second 1.00

Others first, Self second

0.50

0.00 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Discount Factor (!)

Ratio of improvementsocial+product to improvementsocial

1.20

1.00

0.80

0.60 Self first, Others second 0.40

Others first, Self second

0.20

0.00 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Discount Factor (!)

The ratio of improvementsocial to improvementproduct is consistently greater than 1, suggesting that social seed characteristics are more useful than product-related seed characteristics for identifying groups of seeds with higher activity. The ratio of improvementsocial+product to improvementsocial is consistently close to 1, suggesting that product-related seed characteristics do not add much information above and beyond social seed characteristics.

28 FIGURE 3 ROBUSTNESS OF THE RESULTS TO MODEL SPECIFICATION – CAMPAIGN 2

3.00 Self first, Others second Ratio of improvementsocial to improvementproduct

2.50

Others first, Self second

2.00

1.50

1.00

0.50

0.00 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Discount Factor (!)

Ratio of improvementsocial+product to improvementsocial

1.20

1.00

0.80 Self first, Others second

0.60

Others first, Self second 0.40

0.20

0.00 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Discount Factor (!)

The ratio of improvementsocial to improvementproduct is consistently greater than 1, suggesting that social seed characteristics are more useful than product-related seed characteristics for identifying groups of seeds with higher activity. The ratio of improvementsocial+product to improvementsocial is consistently close to 1, suggesting that product-related seed characteristics do not add much information above and beyond social seed characteristics.

29 FIGURE 4 CAMPAIGN 2 COMPARISONS WITH SOCIAL SCORE BASED ON CAMPAIGN 1 ESTIMATES

2.50

Ratio of improvementsocial to improvementproduct

2.00

1.50 Self first, Others second 1.00 Others first, Self second 0.50

0.00 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Discount Factor (!)

The ratio of improvementsocial to improvementproduct is consistently greater than 1, suggesting that social seed characteristics combined with parameter estimates from a previous campaign are more useful in identifying groups of seeds with higher activity as compared to productrelated seed characteristics combined with data from the target campaign.

1

30 TABLE 1 MEASURED SOCIAL AND PRODUCT-RELATED SEED CHARACTERISTICS – CAMPAIGN 1 Question

Response Menu

Social Seed Characteristics “About how many people do you talk to on a daily basis?” “Which best describes you socially?”

“When you recommend products or services to friends, people usually?”

1-9 10-19 20-29 30+ Independent: mostly like to do my own thing Somewhat Social: sometimes I'm social, sometimes not Outgoing: like to socialize but enjoy time to myself Very social: I like to be connected with other people most of the time Not sure Go find out more about the product and determine if they need/want it Go out and buy the product

Product-related Seed Characteristics “How often to you wear nail polish?” (category usage frequency) “How familiar are you with the nail polish brand OPI?” (brand familiarity)

5-point scale (1=Never, 5=All of the time)

“What is your opinion of OPI nail polish?” (brand liking)

5-point scale (1=Very unfavorable, 5=Very favorable)

“How familiar are you with Nic's sticks from OPI?” (product familiarity)

5-point scale (1=Not at all familiar, 5= Very familiar)

“How likely are you to purchase Nic's Sticks from Nicole by OPI?” (product purchase likelihood) “How likely are you to recommend Nic's Sticks from Nicole by OPI?” (product recommendation likelihood)

10-point scale (1=Very unlikely, 5= Neither likely nor unlikely / not familiar with Nic's sticks, 10=Very likely)

5-point scale (1=Not at all familiar, 5=Very familiar)

10-point scale (1=Very unlikely, 5= Neither likely nor unlikely / not familiar with Nic's sticks, 10=Very likely)

31 TABLE 2 PARAMETER ESTIMATES FOR BASE MODEL – CAMPAIGN 1 Covariate

βself Estimate

βother t-value

Estimate

t-value Demographics Age: Below 25 (baseline) — — — — 25 to 29 -.084 -.391 .230** 3.048 30 to 34 -.044 -.203 .141 1.876 35 to 39 .600* 2.382 .380** 4.182 40 to 44 .321 1.205 .220* 2.360 45 to 49 .449 1.465 .254* 2.370 50 and older -.218 -.647 .163 1.402 Employment status * -.215 -1.473 -.123 -2.359 (1 = employed full-time) Product-Related Seed Characteristics Category usage frequency -.134 -1.532 -.098** -3.259 Brand familiarity -.050 -.810 -.049* -2.244 Brand liking -.020 -.276 .019 .760 Product familiarity .271** 2.765 .101** 2.764 Product purchase likelihood -.073 -.986 -.069* -2.566 Product recommendation likelihood .055 .762 .056* 2.131 Social Seed Characteristics About how many people do you talk to on a daily basis? 1-9 (baseline) — — — — 10-19 .370* 2.054 .200** 3.334 20-29 .575* 2.572 .363** 4.640 * 30+ .568 2.494 .545** 6.431 Which best describes you socially? Independent (baseline) — — — — Somewhat Social .823 1.371 .457* 2.072 Outgoing .779 1.316 .500* 2.298 Very Social 1.072 1.779 .660** 2.977 When you recommend products or services to friends, people usually? Not sure (baseline) — — — — Go find out more about the product and * .115 .434 .214 2.397 determine if they need/want it * ** Go out and buy the product .561 2.040 .387 4.156 Other Parameters Intercept -.130 -.169 .000† — * γself/other .228 2.385 1.001** 29.407 θself/other .643** 6.140 .815** 5.870 Model Fit -2 Log-Likelihood 5100.955 Mean estimated probability of chosen coupon .131†† allocation * p < .05, ** p < .01. † The estimate of γother is very close to 1. When γother ≈ 1, the parameter θother and the intercept are not well identified separately. Therefore we set the intercept to 0. †† The mean estimated probability of chosen coupon allocation should be compared to the chance probability of 1/21 =.048 (21 possible combinations of {nself, nothers}).

32 TABLE 3 MEAN COUPONS SHARED WITH OTHERS BASED ON SOCIAL, PRODUCT AND PRODUCT+SOCIAL SCORE QUARTILES – CAMPAIGN 1

Score on which quartiles are based social

4th quartile

3rd quartile

2rd quartile

1st quartile

3.65

3.29

2.96

2.29

product

3.31

3.37

2.94

2.57

product+social

3.68

3.38

2.82

2.32

33 TABLE 4 PARAMETER ESTIMATES FOR THE BASE MODEL – CAMPAIGN 2 Covariate

βself Estimate

βother t-value

Estimate

t-value Demographics Age: Below 25 (baseline) — — — — 25 to 29 -.505 -.229 -.443 -.379 30 to 34 -.320 -.145 -.317 -.272 35 to 39 -.123 -.056 -.279 -.239 40 to 44 .039 .017 -.278 -.238 45 to 49 .170 .077 -.232 -.199 50 and older 1.036 .466 .080 .068 Employment status (1 = employed full-time) -.131 -1.131 -.103* -2.225 a Product-Related Seed Characteristics Brand liking .123* 2.437 .034 1.757 Product familiarity .061 1.060 .016 .693 Product purchase likelihood .185* 2.110 .046 1.335 Product recommendation likelihood .144** 3.622 .047** 2.855 Social Seed Characteristics About how many people do you talk to on a daily basis? 1-9 (baseline) — — — — 10-19 .851 1.825 .330* 2.123 20-29 1.222** 2.568 .554** 3.304 * 30+ 1.035 2.189 .491** 3.028 Which best describes you socially? Independent (baseline) — — — — Somewhat Social .937 1.814 .312 1.783 Outgoing .542 1.113 .252 1.565 Very Social 1.099* 2.223 .470** 2.737 When you recommend products or services to friends, people usually? Not sure (baseline) — — — — Go find out more about the product and -.146 -.425 -.141 -1.130 determine if they need/want it Go out and buy the product .251 .729 .082 .652 Other Parameters Intercept -2.009 -.842 .000† — ** γself/other .413 6.572 1.039** 14.782 θself/other .921** 5.633 .707** .488 Model Fit -2 Log-Likelihood 7015.080 Mean estimated probability of chosen coupon .180†† allocation * p < .05, ** p < .01. † The estimate of γother is close to 1. When γother ≈ 1, the parameter θother and the intercept are not well identified separately. Therefore we set the intercept to 0. †† The mean estimated probability of chosen coupon allocation should be compared to the chance probability of 1/21 =.048 (21 possible combinations of {nself, nothers}).

34 TABLE 5 MEAN COUPONS SHARED WITH OTHERS BASED ON SOCIAL, PRODUCT AND PRODUCT+SOCIAL SCORE QUARTILES – CAMPAIGN 2 Score on which quartiles are based social

4th quartile

3rd quartile

2rd quartile

1st quartile

3.76

3.63

3.45

3.13

product

3.68

3.53

3.39

3.39

product+social

3.77

3.60

3.50

3.11

social (based on Campaign 1 estimates)

3.79

3.60

3.36

3.23

35 APPENDIX 1 ALTERNATIVE DECISION PROCESS

The model described in the paper assumes that seed i first decides how many coupons to use for herself (period 1) and then decides how many coupons to share with others (period 2). Here we describe an alternative model in which the seed first decides how many coupons to share with others (period 1) and then decides how many coupons to use for herself (period 2). Both models are identical when the discount factor δ is equal to 1, in which case a joint decision process is assumed. This alternative model assumes the same utility function specifications as the model described in the paper. Seed i is now assumed to solve the following optimization problems in sequence: Decision 1 : nohers,i = argmax[uothers,i ( nothers ) − cothers,i ( nothers ) + ε1 ( nothers ) + δV2 ( nothers )] nothers

subject to : 0 ≤ nothers ≤ 5 Decision 2 : nself ,i = argmax[uself ,i ( nself ) − cself ,i ( nself ) + ε 2 ( nself )] nself

subject to : 0 ≤ nself ≤ 5 − nothers,i

where V2 (nothers ) = log(

5 − nothers

∑ exp[u

self , i

(nself ) − cself ,i (nself )])

n self = 0

The likelihood function is as follows: Pr({n self ,i , nothers,i }) = P( nself ,i | nothers,i ) ⋅ P( nothers,i ) = exp[u self ,i ( n self ,i ) − cself ,i ( n self ,i )] 5−nothers ,i

∑ exp[u

self ,i ( n self

) − cself ,i ( n self )

⋅

]

nself =0

5− nothers ,i

exp[uothers,i ( nothers,i ) − cothers,i ( nothers,i ) + δ log(

∑ exp[u

self ,i ( n self

) − cself ,i ( nself )])]

nself =0

5−nothers ⎡ ⎤ exp⎢uothers,i ( nothers ) − cothers,i ( nothers ) + δ log( exp[u self ,i ( nself ) − cself ,i ( n self )]) ⎥ ⎢⎣ ⎥⎦ nself =0 nothers =0 5

∑

∑

36 APPENDIX 2 BASE MODEL AS A SPECIAL CASE OF DYNAMIC MODELS

The base model in which seed i jointly decides on the number of coupons used for herself and shared with others is a special case of each dynamic model with δ = 1. Indeed, when δ = 1 the likelihood function in Equation (4) is the same as in Equation (3): Pr({n self ,i , nothers,i }) = P( nothers,i | n self ,i ) ⋅ P( n self ,i ) = exp[uothers,i ( nothers,i ) − cothers,i ( nothers,i )] 5−nself ,i

∑ exp[u

others,i ( nothers ) − cothers,i ( nothers )

⋅

]

nothers ,i =0

exp[u self ,i ( n self ,i ) − cself ,i ( n self ,i ) + log(

5−nselfi

∑ exp[u

others,i ( nothers ) − cothers,i ( nothers )])]

nothers =0

5−nself ⎡ ⎤ exp ⎢u self ,i ( n self ) − cself ,i ( n self ) + log( exp[uothers,i ( nothers ) − cothers,i ( nothers )]) ⎥ ⎢⎣ ⎥⎦ nothers =0 nself ,i =0 5

∑

=

∑

exp[uothers,i (nothers,i ) ! cothers,i (nothers,i )]

'

5!nself ,i

exp "#uothers,i (nothers ) ! cothers,i (nothers )

%$& nothers =0

5!nselfi

exp[uself ,i (nself ,i ) ! cself ,i (nself ,i )]' ( 5

& exp[u

self ,i

exp[uothers,i (nothers ) ! cothers,i (nothers )])

(nself ) ! cself ,i (nself )]' (

&

exp[uothers,i (nothers ) ! cothers,i (nothers )])

nothers =0

nself =0

=

&

nothers =0 5!nself

exp[uothers,i (nothers,i ) ! cothers,i (nothers,i )]" exp[uself ,i (nself ,i ) ! cself ,i (nself ,i )] 5

5!nself

# #

exp[uself ,i (nself ) ! cself ,i (nself )]" exp[uothers,i (nothers ) ! cothers,i (nothers )]

nself =0 nothers =0

=

exp[uothers,i (nothers,i ) ! cothers,i (nothers,i ) + uself ,i (nself ,i ) ! cself ,i (nself ,i )] 5

5!nself

# # nself =0 nothers =0

exp[uself ,i (nself ) ! cself ,i (nself ) + uothers,i (nothers ) ! cothers,i (nothers )]